{"version":2,"relationshipURLBase":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p58.html","relationshipResolution":"Resolve relative relationship URLs against relationshipURLBase; complete records and absolute URLs are in blocks.","title":"用文字组织类别，而不是重做一张固定标签表","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p58.html","markdown":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p58.md","pages":[{"id":"clip-candidate-interface","title":"候选改变时，分数回答的问题也变了","unit":"已知图文配对怎样变成可用的文字候选接口？","unitId":"clip-pairing-to-candidates","chapter":"III","number":1},{"id":"clip-lab04-candidate-controls","title":"在真实猫图上核对变化的量","unit":"已知图文配对怎样变成可用的文字候选接口？","unitId":"clip-pairing-to-candidates","chapter":"III","number":2}],"aliases":[{"oldId":"explanation","targetId":"clip-candidate-interface","operation":"split-explanation-range"},{"oldId":"b-814e5049acf005613ef77f45","targetId":"b-814e5049acf005613ef77f45","operation":"historical-explanation","sourceScope":"clip-lab04-candidate-controls"},{"oldId":"b-29a7e6bb39eb4195d7245d26","targetId":"b-29a7e6bb39eb4195d7245d26","operation":"historical-explanation","sourceScope":"document"},{"oldId":"b-dcbeb4e76d99d8d434c355dc","targetId":"b-dcbeb4e76d99d8d434c355dc","operation":"historical-explanation","sourceScope":"clip-candidate-interface"},{"oldId":"b-b28db2557bf7e7aea691d384","targetId":"b-b28db2557bf7e7aea691d384","page":"clip-candidate-interface","operation":"move","beforeSHA256":"d71cf15e707c8a3674b4c90a599efee783cb84da5b8c8236a1d01c3add7b4958","afterSHA256":"d71cf15e707c8a3674b4c90a599efee783cb84da5b8c8236a1d01c3add7b4958","reason":"Same source body in authored semantic page"},{"oldId":"b-7ba3f0ffc3c2b3e542357d7a","targetId":"b-7ba3f0ffc3c2b3e542357d7a","page":"clip-candidate-interface","operation":"move","beforeSHA256":"c0869b107d85c62b018d7fb65374756604d6b2b1bfb6f21d0634f9d64fc224ae","afterSHA256":"c0869b107d85c62b018d7fb65374756604d6b2b1bfb6f21d0634f9d64fc224ae","reason":"Same source body in authored semantic page"},{"oldId":"b-f12251795d30c9073b8e2454","targetId":"b-f12251795d30c9073b8e2454","page":"clip-candidate-interface","operation":"move","beforeSHA256":"b9a790420611560584ad9070f7e9e60c9da1afe63b7e0688164e330d859589ab","afterSHA256":"b9a790420611560584ad9070f7e9e60c9da1afe63b7e0688164e330d859589ab","reason":"Same source body in authored semantic page"},{"oldId":"b-ca8bf5f96c0df1f8d1cbe32a","targetId":"b-ca8bf5f96c0df1f8d1cbe32a","page":"clip-candidate-interface","operation":"rewrite","beforeSHA256":"efd04c5017e4b87beb89d9f794f9ada7ca64a34203b4be504555b65e22e2e7ed","afterSHA256":"0d2ff1267ce69ac4b9cdb47710781d65f50866ed4dcb2532e838a0973fc39dd7","reason":"Same source identity with figure semantics or reading context enriched"},{"oldId":"b-b08b9dc8b87937086631a1aa","targetId":"b-b08b9dc8b87937086631a1aa","page":"clip-candidate-interface","operation":"move","beforeSHA256":"786ce10a6b32a492259a3dccd0a54e61cd41d28a34382fb7b5386c5f03b7ad69","afterSHA256":"786ce10a6b32a492259a3dccd0a54e61cd41d28a34382fb7b5386c5f03b7ad69","reason":"Same source body in authored semantic page"},{"oldId":"b-83c4522a90c03d32385dc5bf","targetId":"b-83c4522a90c03d32385dc5bf","page":"clip-candidate-interface","operation":"move","beforeSHA256":"b09a39f811e07fced38d135382e298fc60d68f057e8d11bca878aa72275461af","afterSHA256":"b09a39f811e07fced38d135382e298fc60d68f057e8d11bca878aa72275461af","reason":"Same source body in authored semantic page"},{"oldId":"b-27429ee70bf953d480785f8a","targetId":"b-27429ee70bf953d480785f8a","page":"clip-candidate-interface","operation":"move","beforeSHA256":"26a7c9505fa6a166a87fc881d455d4b6f2736e49db95462bd202b20518cb5de4","afterSHA256":"26a7c9505fa6a166a87fc881d455d4b6f2736e49db95462bd202b20518cb5de4","reason":"Same source body in authored semantic page"},{"oldId":"b-1a5d08f27c88addad9d6dc40","targetId":"b-1a5d08f27c88addad9d6dc40","page":"clip-candidate-interface","operation":"move","beforeSHA256":"086e35d8c6f13ba74f478977604e7c9e780a177b02f3aa79c3a25292cdea98f4","afterSHA256":"086e35d8c6f13ba74f478977604e7c9e780a177b02f3aa79c3a25292cdea98f4","reason":"Same source body in authored semantic page"},{"oldId":"b-c74fccf3cad771450b843cf4","targetId":"b-c74fccf3cad771450b843cf4","page":"clip-candidate-interface","operation":"move","beforeSHA256":"e92ba23916b55714915a6ebd28f56b3a4572a9b5437ecfffdd9a435b2e49669d","afterSHA256":"e92ba23916b55714915a6ebd28f56b3a4572a9b5437ecfffdd9a435b2e49669d","reason":"Same source body in authored semantic page"},{"oldId":"b-0a5c51dff517749727786372","targetId":"b-0a5c51dff517749727786372","page":"clip-candidate-interface","operation":"move","beforeSHA256":"6a6219f9c82279ead960e1642fdd534942d4593d554e02e087ac4e57abb0b4de","afterSHA256":"6a6219f9c82279ead960e1642fdd534942d4593d554e02e087ac4e57abb0b4de","reason":"Same source body in authored semantic page"},{"oldId":"b-626d2acb4a6141df1154199c","targetId":"b-626d2acb4a6141df1154199c","page":"clip-candidate-interface","operation":"move","beforeSHA256":"f0fb65af26e6d9ecb0b64bbf069740b4dc2ee1d8c31215d5a62337aac21593c3","afterSHA256":"f0fb65af26e6d9ecb0b64bbf069740b4dc2ee1d8c31215d5a62337aac21593c3","reason":"Same source body in authored semantic page"},{"oldId":"b-5890b16052eecbfb038183d1","targetId":"b-5890b16052eecbfb038183d1","page":"clip-candidate-interface","operation":"move","beforeSHA256":"809c6c9ac3a9f62860dc72ba11a2b15cee1c96eae4c6947d10dda8d40d00e687","afterSHA256":"809c6c9ac3a9f62860dc72ba11a2b15cee1c96eae4c6947d10dda8d40d00e687","reason":"Same source body in authored semantic page"},{"oldId":"b-c55b13dc35e0e9b410f8b7aa","targetId":"b-c55b13dc35e0e9b410f8b7aa","page":"clip-candidate-interface","operation":"move","beforeSHA256":"e8fae5c007071741620c2dc805f4eb7da08e3c7af13bdc31da71fab96a831212","afterSHA256":"e8fae5c007071741620c2dc805f4eb7da08e3c7af13bdc31da71fab96a831212","reason":"Same source body in authored semantic page"},{"oldId":"b-c2872725eba7e944d5fefea1","targetId":"b-c2872725eba7e944d5fefea1","page":"clip-candidate-interface","operation":"move","beforeSHA256":"cdb1479f03cc11bd3e623d9342601fa3381492457d4c8c07d085cb2cc0d26923","afterSHA256":"cdb1479f03cc11bd3e623d9342601fa3381492457d4c8c07d085cb2cc0d26923","reason":"Same source body in authored semantic page"},{"oldId":"b-9cd1e1c79c71805da7c519cf","targetId":"b-9cd1e1c79c71805da7c519cf","page":"clip-candidate-interface","operation":"move","beforeSHA256":"b493e1b4a74e9c4d1783e46a4cb1eedb119a9a6732f88f8d93fa34c1e1f125e3","afterSHA256":"b493e1b4a74e9c4d1783e46a4cb1eedb119a9a6732f88f8d93fa34c1e1f125e3","reason":"Same source body in authored semantic page"},{"oldId":"b-d8de51f1feefefa81c5f2538","targetId":"b-d8de51f1feefefa81c5f2538","page":"clip-candidate-interface","operation":"move","beforeSHA256":"a3b84614e2519618b7bb496b72bde7ddcc8d2ed642302e51f58dd439330916a5","afterSHA256":"a3b84614e2519618b7bb496b72bde7ddcc8d2ed642302e51f58dd439330916a5","reason":"Same source body in authored semantic page"},{"oldId":"b-054955a55efd4f0e9f1682dc","targetId":"b-054955a55efd4f0e9f1682dc","page":"clip-lab04-candidate-controls","operation":"move","beforeSHA256":"8ef459f480d8633165535c202fae7d876f4ff69a4a5b1505c6fdca66a6f48ffb","afterSHA256":"8ef459f480d8633165535c202fae7d876f4ff69a4a5b1505c6fdca66a6f48ffb","reason":"Same source body in authored semantic page"},{"oldId":"b-39665916cb74ce1310551303","targetId":"b-39665916cb74ce1310551303","page":"clip-lab04-candidate-controls","operation":"move","beforeSHA256":"08592cac5e85947afd40374d30d6030489704a2787422ccca7752470cb945f62","afterSHA256":"08592cac5e85947afd40374d30d6030489704a2787422ccca7752470cb945f62","reason":"Same source body in authored semantic page"},{"oldId":"b-f7420cc532f6a99410b5a9ae","targetId":"b-f7420cc532f6a99410b5a9ae","page":"clip-lab04-candidate-controls","operation":"move","beforeSHA256":"c7061a8ad0794772fc5fd4b5f0bdb6e66e098c0aaa2732e4c041ff8bb739810e","afterSHA256":"c7061a8ad0794772fc5fd4b5f0bdb6e66e098c0aaa2732e4c041ff8bb739810e","reason":"Same source body in authored semantic page"},{"oldId":"b-59bc4d9eddd71dc79ae1aed3","targetId":"b-59bc4d9eddd71dc79ae1aed3","page":"clip-lab04-candidate-controls","operation":"move","beforeSHA256":"1fb9bd54f6a6d811bd1b68919726fd1177e6a5264ddb1afa54180876084827ba","afterSHA256":"1fb9bd54f6a6d811bd1b68919726fd1177e6a5264ddb1afa54180876084827ba","reason":"Same source body in authored semantic page"},{"oldId":"b-64711d8f36a2686470a84ea7","targetId":"b-64711d8f36a2686470a84ea7","page":"clip-lab04-candidate-controls","operation":"move","beforeSHA256":"3fd55da4e643785a3ea3b5949b03d73e798ccf3ae1b5812997a45c182c9d7898","afterSHA256":"3fd55da4e643785a3ea3b5949b03d73e798ccf3ae1b5812997a45c182c9d7898","reason":"Same source body in authored semantic page"},{"oldId":"b-48a018c1cdfb14b237f9b6cb","targetId":"b-48a018c1cdfb14b237f9b6cb","page":"clip-lab04-candidate-controls","operation":"move","beforeSHA256":"f16be0424762a44344338696f06ef266c4571d15367ae2cbbd28fac98ac7e9cb","afterSHA256":"f16be0424762a44344338696f06ef266c4571d15367ae2cbbd28fac98ac7e9cb","reason":"Same source body in authored semantic page"},{"oldId":"b-feec39ed6d00fa6ad145a1f8","targetId":"b-feec39ed6d00fa6ad145a1f8","page":"clip-lab04-candidate-controls","operation":"move","beforeSHA256":"b25685f04cd1b885f3a5999e6a912460c7424021cdc5b2190584242086a82681","afterSHA256":"b25685f04cd1b885f3a5999e6a912460c7424021cdc5b2190584242086a82681","reason":"Same source body in authored semantic page"},{"oldId":"b-68dac1c085af6e65f27c5515","targetId":"b-68dac1c085af6e65f27c5515","page":"clip-lab04-candidate-controls","operation":"move","beforeSHA256":"491fc9e14f1912ea4bbbc4e9d1c5a17b30cd6ae22d093d54b6b170c9432077c8","afterSHA256":"491fc9e14f1912ea4bbbc4e9d1c5a17b30cd6ae22d093d54b6b170c9432077c8","reason":"Same source body in authored semantic page"},{"oldId":"b-6613a34dadedb0edcb265e6a","targetId":"b-6613a34dadedb0edcb265e6a","page":"clip-lab04-candidate-controls","operation":"move","beforeSHA256":"335d4c8c80f3ea1e4a56444e3d790a917f3d860fb0499b78a839f28a7ab795d6","afterSHA256":"335d4c8c80f3ea1e4a56444e3d790a917f3d860fb0499b78a839f28a7ab795d6","reason":"Same source body in authored semantic page"},{"oldId":"b-ad2aabd871e0840c1a9b6d09","targetId":"b-ad2aabd871e0840c1a9b6d09","page":"clip-lab04-candidate-controls","operation":"move","beforeSHA256":"cf44fcc62782fb4f6bdc42c16c5187d57556a03cb7ce6065e7bd9c36d8e7d733","afterSHA256":"cf44fcc62782fb4f6bdc42c16c5187d57556a03cb7ce6065e7bd9c36d8e7d733","reason":"Same source body in authored semantic page"},{"oldId":"b-196ea3d15b85ad404d6ec034","targetId":"b-196ea3d15b85ad404d6ec034","page":"clip-lab04-candidate-controls","operation":"move","beforeSHA256":"b8695bb1b1bba8ad0c7695972964a508c00ec40840e1197f42629041cad9e7e4","afterSHA256":"b8695bb1b1bba8ad0c7695972964a508c00ec40840e1197f42629041cad9e7e4","reason":"Same source body in authored semantic page"},{"oldId":"clip-candidate-interface-title","targetId":"b-876393db9d52e2cfbd798a0a","operation":"retain-source-identity"},{"oldId":"clip-lab04-candidate-controls-title","targetId":"b-692ae89a502f0bf6badb092c","operation":"retain-source-identity"},{"oldId":"clip-lab04-candidate-controls-bridge","targetId":"b-24efd7fa00af8c66e1e50a9a","operation":"retain-source-identity"},{"oldId":"clip-lab04-saved-figure","targetId":"b-94fb66967ed81eea63fd9be1","operation":"retain-source-identity"}],"blocks":[{"id":"b-876393db9d52e2cfbd798a0a","type":"h2","scope":"clip-candidate-interface","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p58.html#b-876393db9d52e2cfbd798a0a","content":"## 候选改变时，分数回答的问题也变了","sha256":"977d6b90c858b99773a18855e89bbbc32dfa5fbb595d33bc952775cb46da7b0e","chapter":{"id":"III","title":"视觉与语言"},"unit":{"id":"clip-pairing-to-candidates","title":"已知图文配对怎样变成可用的文字候选接口？"},"page":{"id":"clip-candidate-interface","title":"候选改变时，分数回答的问题也变了","number":1},"sectionTitle":"候选改变时，分数回答的问题也变了","currentQuestion":"从一对输入到双向损失，再到真实候选干预，哪些量由训练决定、哪些由使用者决定？","sourceSlides":["l3-r2-p58"],"knowledgeBlock":null,"formulas":[],"previousBlock":null,"nextBlock":"b-b28db2557bf7e7aea691d384","prerequisites":[{"id":"b-f12251795d30c9073b8e2454","url":"#b-f12251795d30c9073b8e2454","role":"problem-context"}],"relatedBlocks":[{"id":"b-ca8bf5f96c0df1f8d1cbe32a","url":"#b-ca8bf5f96c0df1f8d1cbe32a","relation":"figure-and-explanation"}]},{"id":"b-b28db2557bf7e7aea691d384","type":"h2","scope":"explanation","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p58.html#b-b28db2557bf7e7aea691d384","content":"## 解释、算例与证据","sha256":"d71cf15e707c8a3674b4c90a599efee783cb84da5b8c8236a1d01c3add7b4958","chapter":{"id":"III","title":"视觉与语言"},"unit":{"id":"clip-pairing-to-candidates","title":"已知图文配对怎样变成可用的文字候选接口？"},"page":{"id":"clip-candidate-interface","title":"候选改变时，分数回答的问题也变了","number":1},"sectionTitle":"候选改变时，分数回答的问题也变了","currentQuestion":"从一对输入到双向损失，再到真实候选干预，哪些量由训练决定、哪些由使用者决定？","sourceSlides":["l3-r2-p58"],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-876393db9d52e2cfbd798a0a","nextBlock":"b-7ba3f0ffc3c2b3e542357d7a","prerequisites":[{"id":"b-876393db9d52e2cfbd798a0a","url":"#b-876393db9d52e2cfbd798a0a","role":"section-title"},{"id":"b-f12251795d30c9073b8e2454","url":"#b-f12251795d30c9073b8e2454","role":"problem-context"}],"relatedBlocks":[{"id":"b-ca8bf5f96c0df1f8d1cbe32a","url":"#b-ca8bf5f96c0df1f8d1cbe32a","relation":"figure-and-explanation"}]},{"id":"b-7ba3f0ffc3c2b3e542357d7a","type":"h1","scope":"document","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p58.html#b-7ba3f0ffc3c2b3e542357d7a","content":"# P58 · 用文字组织类别，而不是重做一张固定标签表","sha256":"c0869b107d85c62b018d7fb65374756604d6b2b1bfb6f21d0634f9d64fc224ae","chapter":{"id":"III","title":"视觉与语言"},"unit":{"id":"clip-pairing-to-candidates","title":"已知图文配对怎样变成可用的文字候选接口？"},"page":{"id":"clip-candidate-interface","title":"候选改变时，分数回答的问题也变了","number":1},"sectionTitle":"候选改变时，分数回答的问题也变了","currentQuestion":"从一对输入到双向损失，再到真实候选干预，哪些量由训练决定、哪些由使用者决定？","sourceSlides":["l3-r2-p58"],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-b28db2557bf7e7aea691d384","nextBlock":"b-f12251795d30c9073b8e2454","prerequisites":[{"id":"b-b28db2557bf7e7aea691d384","url":"#b-b28db2557bf7e7aea691d384","role":"section-title"},{"id":"b-f12251795d30c9073b8e2454","url":"#b-f12251795d30c9073b8e2454","role":"problem-context"}],"relatedBlocks":[{"id":"b-ca8bf5f96c0df1f8d1cbe32a","url":"#b-ca8bf5f96c0df1f8d1cbe32a","relation":"figure-and-explanation"}]},{"id":"b-f12251795d30c9073b8e2454","type":"p","scope":"document","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p58.html#b-f12251795d30c9073b8e2454","content":"这里的 zero-shot 不使用当前下游数据集再训练模型，而用已有编码器比较文字候选。增加候选和更换模板是不同干预：前者可以只变 softmax 分母，后者会改变文本向量。LAB04 的真实猫图四臂比较保留，可直接阅读已有结果，不必为理解接口重跑模型。","sha256":"b9a790420611560584ad9070f7e9e60c9da1afe63b7e0688164e330d859589ab","chapter":{"id":"III","title":"视觉与语言"},"unit":{"id":"clip-pairing-to-candidates","title":"已知图文配对怎样变成可用的文字候选接口？"},"page":{"id":"clip-candidate-interface","title":"候选改变时，分数回答的问题也变了","number":1},"sectionTitle":"候选改变时，分数回答的问题也变了","currentQuestion":"从一对输入到双向损失，再到真实候选干预，哪些量由训练决定、哪些由使用者决定？","sourceSlides":["l3-r2-p58"],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-7ba3f0ffc3c2b3e542357d7a","nextBlock":"b-ca8bf5f96c0df1f8d1cbe32a","prerequisites":[{"id":"b-7ba3f0ffc3c2b3e542357d7a","url":"#b-7ba3f0ffc3c2b3e542357d7a","role":"section-title"}],"relatedBlocks":[{"id":"b-ca8bf5f96c0df1f8d1cbe32a","url":"#b-ca8bf5f96c0df1f8d1cbe32a","relation":"figure-and-explanation"}]},{"id":"b-ca8bf5f96c0df1f8d1cbe32a","type":"figure","scope":"document","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p58.html#b-ca8bf5f96c0df1f8d1cbe32a","content":"图内标签与数值：图像编码器；红色杯子；蓝色杯子；红色球；红色马克杯；同一图像表示；增加近义候选，归一化分数也会改变。；zero-shot：不使用当前下游数据集再训练模型；候选与连线为使用接口示意；不填写未运行的候选分数，也不声称概念从未出现在预训练中。\n![固定的红杯图像](/assets/display/9a66faeae8dd2c354bc56482.webp)\n固定图像编码；文字候选依次为红色杯子、蓝色杯子、红色球，再加入红色马克杯。候选与连线仅示意接口，未标出未运行的分数。\n同一红杯图像编码保持不变，候选从红杯、蓝杯、红球三项增加近义的红色马克杯。 [查看对应图解状态](/#/l3-r2-p58/1)[查看完整图解](/course/figures/l3-r2-p58-1.html)","sha256":"0d2ff1267ce69ac4b9cdb47710781d65f50866ed4dcb2532e838a0973fc39dd7","chapter":{"id":"III","title":"视觉与语言"},"unit":{"id":"clip-pairing-to-candidates","title":"已知图文配对怎样变成可用的文字候选接口？"},"page":{"id":"clip-candidate-interface","title":"候选改变时，分数回答的问题也变了","number":1},"sectionTitle":"候选改变时，分数回答的问题也变了","currentQuestion":"从一对输入到双向损失，再到真实候选干预，哪些量由训练决定、哪些由使用者决定？","sourceSlides":["l3-r2-p58"],"knowledgeBlock":null,"formulas":[],"figure":{"caption":"同一红杯图像编码保持不变，候选从红杯、蓝杯、红球三项增加近义的红色马克杯。 查看对应图解状态","labels":["图像编码器","红色杯子","蓝色杯子","红色球","红色马克杯","同一图像表示","增加近义候选，归一化分数也会改变。","zero-shot：不使用当前下游数据集再训练模型","候选与连线为使用接口示意；不填写未运行的候选分数，也不声称概念从未出现在预训练中。"],"description":"固定图像编码；文字候选依次为红色杯子、蓝色杯子、红色球，再加入红色马克杯。候选与连线仅示意接口，未标出未运行的分数。","evidenceIdentity":"同一红杯图像编码保持不变，候选从红杯、蓝杯、红球三项增加近义的红色马克杯。 查看对应图解状态","images":[{"src":"/assets/display/9a66faeae8dd2c354bc56482.webp","alt":""},{"src":"/assets/display/9a66faeae8dd2c354bc56482.webp","alt":"固定的红杯图像"}]},"previousBlock":"b-f12251795d30c9073b8e2454","nextBlock":"b-b08b9dc8b87937086631a1aa","prerequisites":[{"id":"b-7ba3f0ffc3c2b3e542357d7a","url":"#b-7ba3f0ffc3c2b3e542357d7a","role":"section-title"},{"id":"b-f12251795d30c9073b8e2454","url":"#b-f12251795d30c9073b8e2454","role":"problem-context"}],"relatedBlocks":[]},{"id":"b-b08b9dc8b87937086631a1aa","type":"h2","scope":"narrative","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p58.html#b-b08b9dc8b87937086631a1aa","content":"## 课堂讲述","sha256":"786ce10a6b32a492259a3dccd0a54e61cd41d28a34382fb7b5386c5f03b7ad69","chapter":{"id":"III","title":"视觉与语言"},"unit":{"id":"clip-pairing-to-candidates","title":"已知图文配对怎样变成可用的文字候选接口？"},"page":{"id":"clip-candidate-interface","title":"候选改变时，分数回答的问题也变了","number":1},"sectionTitle":"候选改变时，分数回答的问题也变了","currentQuestion":"从一对输入到双向损失，再到真实候选干预，哪些量由训练决定、哪些由使用者决定？","sourceSlides":["l3-r2-p58"],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-ca8bf5f96c0df1f8d1cbe32a","nextBlock":"b-83c4522a90c03d32385dc5bf","prerequisites":[{"id":"b-7ba3f0ffc3c2b3e542357d7a","url":"#b-7ba3f0ffc3c2b3e542357d7a","role":"section-title"},{"id":"b-f12251795d30c9073b8e2454","url":"#b-f12251795d30c9073b8e2454","role":"problem-context"}],"relatedBlocks":[{"id":"b-ca8bf5f96c0df1f8d1cbe32a","url":"#b-ca8bf5f96c0df1f8d1cbe32a","relation":"figure-and-explanation"}]},{"id":"b-83c4522a90c03d32385dc5bf","type":"p","scope":"narrative","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p58.html#b-83c4522a90c03d32385dc5bf","content":"传统分类头的输出位置绑定训练时的类别集合。CLIP 可以把我们提供的类别描述先编码，再与图像比较，这就让使用接口更灵活。\n这里的 zero-shot 用法，是不使用当前下游数据集再训练模型，直接用已有的图像和文字编码器比较候选描述。它不保证相关概念从未出现在预训练资料中。改变候选文字，会改变这次比较的答案集合；加入新候选会改变 softmax 归一化分数，因此这些分数不能直接视为稳定校准的真实概率。\n最重要的是看：相同视觉表示怎样在文字描述的帮助下进入一个新任务。","sha256":"b09a39f811e07fced38d135382e298fc60d68f057e8d11bca878aa72275461af","chapter":{"id":"III","title":"视觉与语言"},"unit":{"id":"clip-pairing-to-candidates","title":"已知图文配对怎样变成可用的文字候选接口？"},"page":{"id":"clip-candidate-interface","title":"候选改变时，分数回答的问题也变了","number":1},"sectionTitle":"候选改变时，分数回答的问题也变了","currentQuestion":"从一对输入到双向损失，再到真实候选干预，哪些量由训练决定、哪些由使用者决定？","sourceSlides":["l3-r2-p58"],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-b08b9dc8b87937086631a1aa","nextBlock":"b-27429ee70bf953d480785f8a","prerequisites":[{"id":"b-b08b9dc8b87937086631a1aa","url":"#b-b08b9dc8b87937086631a1aa","role":"section-title"},{"id":"b-f12251795d30c9073b8e2454","url":"#b-f12251795d30c9073b8e2454","role":"problem-context"}],"relatedBlocks":[{"id":"b-ca8bf5f96c0df1f8d1cbe32a","url":"#b-ca8bf5f96c0df1f8d1cbe32a","relation":"figure-and-explanation"}]},{"id":"b-27429ee70bf953d480785f8a","type":"h3","scope":"clip-candidate-probability","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p58.html#b-27429ee70bf953d480785f8a","content":"### 换候选、换模板、换图像：三个不同干预","sha256":"26a7c9505fa6a166a87fc881d455d4b6f2736e49db95462bd202b20518cb5de4","chapter":{"id":"III","title":"视觉与语言"},"unit":{"id":"clip-pairing-to-candidates","title":"已知图文配对怎样变成可用的文字候选接口？"},"page":{"id":"clip-candidate-interface","title":"候选改变时，分数回答的问题也变了","number":1},"sectionTitle":"换候选、换模板、换图像：三个不同干预","currentQuestion":"从一对输入到双向损失，再到真实候选干预，哪些量由训练决定、哪些由使用者决定？","sourceSlides":["l3-r2-p58"],"knowledgeBlock":"clip-candidate-probability","formulas":[],"previousBlock":"b-83c4522a90c03d32385dc5bf","nextBlock":"b-1a5d08f27c88addad9d6dc40","prerequisites":[{"id":"b-b08b9dc8b87937086631a1aa","url":"#b-b08b9dc8b87937086631a1aa","role":"section-title"},{"id":"b-f12251795d30c9073b8e2454","url":"#b-f12251795d30c9073b8e2454","role":"problem-context"}],"relatedBlocks":[{"id":"b-ca8bf5f96c0df1f8d1cbe32a","url":"#b-ca8bf5f96c0df1f8d1cbe32a","relation":"figure-and-explanation"},{"id":"b-1a5d08f27c88addad9d6dc40","url":"#b-1a5d08f27c88addad9d6dc40","relation":"same-reasoning"},{"id":"b-c74fccf3cad771450b843cf4","url":"#b-c74fccf3cad771450b843cf4","relation":"same-reasoning"},{"id":"b-0a5c51dff517749727786372","url":"#b-0a5c51dff517749727786372","relation":"same-reasoning"},{"id":"b-626d2acb4a6141df1154199c","url":"#b-626d2acb4a6141df1154199c","relation":"same-reasoning"}]},{"id":"b-1a5d08f27c88addad9d6dc40","type":"p","scope":"clip-candidate-probability","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p58.html#b-1a5d08f27c88addad9d6dc40","content":"训练完后，CLIP 可以把类别写成文字候选，例如 `a photo of a cat.`，再比较图像与每条候选的表示。这里的 zero-shot 用法，是不使用当前下游数据集再训练模型，直接用已有的图像和文字编码器比较候选描述。它不保证相关概念从未出现在预训练资料中。改变候选文字，会改变这次比较的答案集合；加入新候选会改变 softmax 归一化分数，因此这些分数不能直接视为稳定校准的真实概率。","sha256":"086e35d8c6f13ba74f478977604e7c9e780a177b02f3aa79c3a25292cdea98f4","chapter":{"id":"III","title":"视觉与语言"},"unit":{"id":"clip-pairing-to-candidates","title":"已知图文配对怎样变成可用的文字候选接口？"},"page":{"id":"clip-candidate-interface","title":"候选改变时，分数回答的问题也变了","number":1},"sectionTitle":"换候选、换模板、换图像：三个不同干预","currentQuestion":"从一对输入到双向损失，再到真实候选干预，哪些量由训练决定、哪些由使用者决定？","sourceSlides":["l3-r2-p58"],"knowledgeBlock":"clip-candidate-probability","formulas":[],"previousBlock":"b-27429ee70bf953d480785f8a","nextBlock":"b-c74fccf3cad771450b843cf4","prerequisites":[{"id":"b-27429ee70bf953d480785f8a","url":"#b-27429ee70bf953d480785f8a","role":"section-title"},{"id":"b-f12251795d30c9073b8e2454","url":"#b-f12251795d30c9073b8e2454","role":"problem-context"}],"relatedBlocks":[{"id":"b-ca8bf5f96c0df1f8d1cbe32a","url":"#b-ca8bf5f96c0df1f8d1cbe32a","relation":"figure-and-explanation"},{"id":"b-27429ee70bf953d480785f8a","url":"#b-27429ee70bf953d480785f8a","relation":"same-reasoning"},{"id":"b-c74fccf3cad771450b843cf4","url":"#b-c74fccf3cad771450b843cf4","relation":"same-reasoning"},{"id":"b-0a5c51dff517749727786372","url":"#b-0a5c51dff517749727786372","relation":"same-reasoning"},{"id":"b-626d2acb4a6141df1154199c","url":"#b-626d2acb4a6141df1154199c","relation":"same-reasoning"}]},{"id":"b-c74fccf3cad771450b843cf4","type":"p","scope":"clip-candidate-probability","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p58.html#b-c74fccf3cad771450b843cf4","content":"设猫和狗的 logits 是 2 与 1，候选 softmax 给猫 $e^2/(e^2+e)=0.7311$。新增一个 logit 同为 2 的幼猫候选，原猫概率变为 $e^2/(2e^2+e)=0.4223$。同一图像、同一猫文字、同一点积都没有改变，只是分母和选项变了。","sha256":"e92ba23916b55714915a6ebd28f56b3a4572a9b5437ecfffdd9a435b2e49669d","chapter":{"id":"III","title":"视觉与语言"},"unit":{"id":"clip-pairing-to-candidates","title":"已知图文配对怎样变成可用的文字候选接口？"},"page":{"id":"clip-candidate-interface","title":"候选改变时，分数回答的问题也变了","number":1},"sectionTitle":"换候选、换模板、换图像：三个不同干预","currentQuestion":"从一对输入到双向损失，再到真实候选干预，哪些量由训练决定、哪些由使用者决定？","sourceSlides":["l3-r2-p58"],"knowledgeBlock":"clip-candidate-probability","formulas":[{"latex":"e^2/(e^2+e)=0.7311","definitionContext":"训练完后，CLIP 可以把类别写成文字候选，例如 `a photo of a cat.`，再比较图像与每条候选的表示。这里的 zero-shot 用法，是不使用当前下游数据集再训练模型，直接用已有的图像和文字编码器比较候选描述。它不保证相关概念从未出现在预训练资料中。改变候选文字，会改变这次比较的答案集合；加入新候选会改变 softmax 归一化分数，因此这些分数不能直接视为稳定校准的真实概率。\n设猫和狗的 logits 是 2 与 1，候选 softmax 给猫 $e^2/(e^2+e)=0.7311$。新增一个 logit 同为 2 的幼猫候选，原猫概率变为 $e^2/(2e^2+e)=0.4223$。同一图像、同一猫文字、同一点积都没有改变，只是分母和选项变了。\n因此有三个需要隔离的实验。只增候选：已有候选余弦应保持，softmax 可以改变。只改文字模板：文字向量改变，余弦也可能改变。只换图像：图像向量改变，应检查哪些候选关系随视觉内容变化。若三者一起改，就难以解释分数差异。\nLAB04 保留的四臂猫图实验专门做这件事：A 是三类基础候选，B 添加 tiger/fox，C 保持三类但改 cropped 模板，D 从 B 删除 cat。删除正确选项后仍然必有最高项；这不是模型发现照片没有猫，而是回答空间已经被使用者改变。"},{"latex":"e^2/(2e^2+e)=0.4223","definitionContext":"训练完后，CLIP 可以把类别写成文字候选，例如 `a photo of a cat.`，再比较图像与每条候选的表示。这里的 zero-shot 用法，是不使用当前下游数据集再训练模型，直接用已有的图像和文字编码器比较候选描述。它不保证相关概念从未出现在预训练资料中。改变候选文字，会改变这次比较的答案集合；加入新候选会改变 softmax 归一化分数，因此这些分数不能直接视为稳定校准的真实概率。\n设猫和狗的 logits 是 2 与 1，候选 softmax 给猫 $e^2/(e^2+e)=0.7311$。新增一个 logit 同为 2 的幼猫候选，原猫概率变为 $e^2/(2e^2+e)=0.4223$。同一图像、同一猫文字、同一点积都没有改变，只是分母和选项变了。\n因此有三个需要隔离的实验。只增候选：已有候选余弦应保持，softmax 可以改变。只改文字模板：文字向量改变，余弦也可能改变。只换图像：图像向量改变，应检查哪些候选关系随视觉内容变化。若三者一起改，就难以解释分数差异。\nLAB04 保留的四臂猫图实验专门做这件事：A 是三类基础候选，B 添加 tiger/fox，C 保持三类但改 cropped 模板，D 从 B 删除 cat。删除正确选项后仍然必有最高项；这不是模型发现照片没有猫，而是回答空间已经被使用者改变。"}],"previousBlock":"b-1a5d08f27c88addad9d6dc40","nextBlock":"b-0a5c51dff517749727786372","prerequisites":[{"id":"b-27429ee70bf953d480785f8a","url":"#b-27429ee70bf953d480785f8a","role":"section-title"},{"id":"b-f12251795d30c9073b8e2454","url":"#b-f12251795d30c9073b8e2454","role":"problem-context"}],"relatedBlocks":[{"id":"b-ca8bf5f96c0df1f8d1cbe32a","url":"#b-ca8bf5f96c0df1f8d1cbe32a","relation":"figure-and-explanation"},{"id":"b-27429ee70bf953d480785f8a","url":"#b-27429ee70bf953d480785f8a","relation":"same-reasoning"},{"id":"b-1a5d08f27c88addad9d6dc40","url":"#b-1a5d08f27c88addad9d6dc40","relation":"same-reasoning"},{"id":"b-0a5c51dff517749727786372","url":"#b-0a5c51dff517749727786372","relation":"same-reasoning"},{"id":"b-626d2acb4a6141df1154199c","url":"#b-626d2acb4a6141df1154199c","relation":"same-reasoning"}]},{"id":"b-0a5c51dff517749727786372","type":"p","scope":"clip-candidate-probability","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p58.html#b-0a5c51dff517749727786372","content":"因此有三个需要隔离的实验。只增候选：已有候选余弦应保持，softmax 可以改变。只改文字模板：文字向量改变，余弦也可能改变。只换图像：图像向量改变，应检查哪些候选关系随视觉内容变化。若三者一起改，就难以解释分数差异。","sha256":"6a6219f9c82279ead960e1642fdd534942d4593d554e02e087ac4e57abb0b4de","chapter":{"id":"III","title":"视觉与语言"},"unit":{"id":"clip-pairing-to-candidates","title":"已知图文配对怎样变成可用的文字候选接口？"},"page":{"id":"clip-candidate-interface","title":"候选改变时，分数回答的问题也变了","number":1},"sectionTitle":"换候选、换模板、换图像：三个不同干预","currentQuestion":"从一对输入到双向损失，再到真实候选干预，哪些量由训练决定、哪些由使用者决定？","sourceSlides":["l3-r2-p58"],"knowledgeBlock":"clip-candidate-probability","formulas":[],"previousBlock":"b-c74fccf3cad771450b843cf4","nextBlock":"b-626d2acb4a6141df1154199c","prerequisites":[{"id":"b-27429ee70bf953d480785f8a","url":"#b-27429ee70bf953d480785f8a","role":"section-title"},{"id":"b-f12251795d30c9073b8e2454","url":"#b-f12251795d30c9073b8e2454","role":"problem-context"}],"relatedBlocks":[{"id":"b-ca8bf5f96c0df1f8d1cbe32a","url":"#b-ca8bf5f96c0df1f8d1cbe32a","relation":"figure-and-explanation"},{"id":"b-27429ee70bf953d480785f8a","url":"#b-27429ee70bf953d480785f8a","relation":"same-reasoning"},{"id":"b-1a5d08f27c88addad9d6dc40","url":"#b-1a5d08f27c88addad9d6dc40","relation":"same-reasoning"},{"id":"b-c74fccf3cad771450b843cf4","url":"#b-c74fccf3cad771450b843cf4","relation":"same-reasoning"},{"id":"b-626d2acb4a6141df1154199c","url":"#b-626d2acb4a6141df1154199c","relation":"same-reasoning"}]},{"id":"b-626d2acb4a6141df1154199c","type":"p","scope":"clip-candidate-probability","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p58.html#b-626d2acb4a6141df1154199c","content":"LAB04 保留的四臂猫图实验专门做这件事：A 是三类基础候选，B 添加 tiger/fox，C 保持三类但改 cropped 模板，D 从 B 删除 cat。删除正确选项后仍然必有最高项；这不是模型发现照片没有猫，而是回答空间已经被使用者改变。","sha256":"f0fb65af26e6d9ecb0b64bbf069740b4dc2ee1d8c31215d5a62337aac21593c3","chapter":{"id":"III","title":"视觉与语言"},"unit":{"id":"clip-pairing-to-candidates","title":"已知图文配对怎样变成可用的文字候选接口？"},"page":{"id":"clip-candidate-interface","title":"候选改变时，分数回答的问题也变了","number":1},"sectionTitle":"换候选、换模板、换图像：三个不同干预","currentQuestion":"从一对输入到双向损失，再到真实候选干预，哪些量由训练决定、哪些由使用者决定？","sourceSlides":["l3-r2-p58"],"knowledgeBlock":"clip-candidate-probability","formulas":[],"previousBlock":"b-0a5c51dff517749727786372","nextBlock":"b-0ada9f2a3352f61b0e1ef614","prerequisites":[{"id":"b-27429ee70bf953d480785f8a","url":"#b-27429ee70bf953d480785f8a","role":"section-title"},{"id":"b-f12251795d30c9073b8e2454","url":"#b-f12251795d30c9073b8e2454","role":"problem-context"}],"relatedBlocks":[{"id":"b-ca8bf5f96c0df1f8d1cbe32a","url":"#b-ca8bf5f96c0df1f8d1cbe32a","relation":"figure-and-explanation"},{"id":"b-27429ee70bf953d480785f8a","url":"#b-27429ee70bf953d480785f8a","relation":"same-reasoning"},{"id":"b-1a5d08f27c88addad9d6dc40","url":"#b-1a5d08f27c88addad9d6dc40","relation":"same-reasoning"},{"id":"b-c74fccf3cad771450b843cf4","url":"#b-c74fccf3cad771450b843cf4","relation":"same-reasoning"},{"id":"b-0a5c51dff517749727786372","url":"#b-0a5c51dff517749727786372","relation":"same-reasoning"}]},{"id":"b-0ada9f2a3352f61b0e1ef614","type":"h4","scope":"clip-candidate-interface","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p58.html#b-0ada9f2a3352f61b0e1ef614","content":"#### 图示条件与公式约定","sha256":"ef7f915c1ae63ed67bd23fecf050249273ba0efa8965d2c8d1c1c6b2d3178122","chapter":{"id":"III","title":"视觉与语言"},"unit":{"id":"clip-pairing-to-candidates","title":"已知图文配对怎样变成可用的文字候选接口？"},"page":{"id":"clip-candidate-interface","title":"候选改变时，分数回答的问题也变了","number":1},"sectionTitle":"候选改变时，分数回答的问题也变了","currentQuestion":"从一对输入到双向损失，再到真实候选干预，哪些量由训练决定、哪些由使用者决定？","sourceSlides":["l3-r2-p58"],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-626d2acb4a6141df1154199c","nextBlock":"b-5890b16052eecbfb038183d1","prerequisites":[{"id":"b-27429ee70bf953d480785f8a","url":"#b-27429ee70bf953d480785f8a","role":"section-title"},{"id":"b-f12251795d30c9073b8e2454","url":"#b-f12251795d30c9073b8e2454","role":"problem-context"}],"relatedBlocks":[{"id":"b-ca8bf5f96c0df1f8d1cbe32a","url":"#b-ca8bf5f96c0df1f8d1cbe32a","relation":"figure-and-explanation"}]},{"id":"b-5890b16052eecbfb038183d1","type":"p","scope":"document","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p58.html#b-5890b16052eecbfb038183d1","content":"候选集由课程明确给定。两个状态共用同一图像表示，第二状态只增加“红色马克杯”文本候选；不提供未运行的模型分数。候选 softmax 随比较集合变化，不等于已校准的开放世界概率。","sha256":"809c6c9ac3a9f62860dc72ba11a2b15cee1c96eae4c6947d10dda8d40d00e687","chapter":{"id":"III","title":"视觉与语言"},"unit":{"id":"clip-pairing-to-candidates","title":"已知图文配对怎样变成可用的文字候选接口？"},"page":{"id":"clip-candidate-interface","title":"候选改变时，分数回答的问题也变了","number":1},"sectionTitle":"候选改变时，分数回答的问题也变了","currentQuestion":"从一对输入到双向损失，再到真实候选干预，哪些量由训练决定、哪些由使用者决定？","sourceSlides":["l3-r2-p58"],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-0ada9f2a3352f61b0e1ef614","nextBlock":"b-b55fb389a499afc0cecef476","prerequisites":[{"id":"b-0ada9f2a3352f61b0e1ef614","url":"#b-0ada9f2a3352f61b0e1ef614","role":"section-title"},{"id":"b-f12251795d30c9073b8e2454","url":"#b-f12251795d30c9073b8e2454","role":"problem-context"}],"relatedBlocks":[{"id":"b-ca8bf5f96c0df1f8d1cbe32a","url":"#b-ca8bf5f96c0df1f8d1cbe32a","relation":"figure-and-explanation"}]},{"id":"b-b55fb389a499afc0cecef476","type":"h4","scope":"retained-reading","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p58.html#b-b55fb389a499afc0cecef476","content":"#### 这条线索的展开与后续阅读","sha256":"48b8229f7559b36ed4d5628416e3991e5c73293fbce43374db67386c8736f3bc","chapter":{"id":"III","title":"视觉与语言"},"unit":{"id":"clip-pairing-to-candidates","title":"已知图文配对怎样变成可用的文字候选接口？"},"page":{"id":"clip-candidate-interface","title":"候选改变时，分数回答的问题也变了","number":1},"sectionTitle":"候选改变时，分数回答的问题也变了","currentQuestion":"从一对输入到双向损失，再到真实候选干预，哪些量由训练决定、哪些由使用者决定？","sourceSlides":["l3-r2-p58"],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-5890b16052eecbfb038183d1","nextBlock":"b-c55b13dc35e0e9b410f8b7aa","prerequisites":[{"id":"b-0ada9f2a3352f61b0e1ef614","url":"#b-0ada9f2a3352f61b0e1ef614","role":"section-title"},{"id":"b-f12251795d30c9073b8e2454","url":"#b-f12251795d30c9073b8e2454","role":"problem-context"}],"relatedBlocks":[{"id":"b-ca8bf5f96c0df1f8d1cbe32a","url":"#b-ca8bf5f96c0df1f8d1cbe32a","relation":"figure-and-explanation"}]},{"id":"b-c55b13dc35e0e9b410f8b7aa","type":"ul","scope":"retained-reading","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p58.html#b-c55b13dc35e0e9b410f8b7aa","content":"- [训练结束后，文字变成可换的候选接口](/supplement.html?return=l3-r2-p58/0#/l3-clip-use/0) 本页说明\n- [CLIP 从图文对走到候选接口｜怎样检查这项解释](/supplement.html?return=l3-r2-p58/0#/l3-clip-full-chain-evidence/0) 本页说明\n- [CLIP 从图文对走到候选接口｜什么结论还不能推出](/supplement.html?return=l3-r2-p58/0#/l3-clip-full-chain-boundary/0) 本页说明","sha256":"e8fae5c007071741620c2dc805f4eb7da08e3c7af13bdc31da71fab96a831212","chapter":{"id":"III","title":"视觉与语言"},"unit":{"id":"clip-pairing-to-candidates","title":"已知图文配对怎样变成可用的文字候选接口？"},"page":{"id":"clip-candidate-interface","title":"候选改变时，分数回答的问题也变了","number":1},"sectionTitle":"候选改变时，分数回答的问题也变了","currentQuestion":"从一对输入到双向损失，再到真实候选干预，哪些量由训练决定、哪些由使用者决定？","sourceSlides":["l3-r2-p58"],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-b55fb389a499afc0cecef476","nextBlock":"b-c2872725eba7e944d5fefea1","prerequisites":[{"id":"b-b55fb389a499afc0cecef476","url":"#b-b55fb389a499afc0cecef476","role":"section-title"},{"id":"b-f12251795d30c9073b8e2454","url":"#b-f12251795d30c9073b8e2454","role":"problem-context"}],"relatedBlocks":[{"id":"b-ca8bf5f96c0df1f8d1cbe32a","url":"#b-ca8bf5f96c0df1f8d1cbe32a","relation":"figure-and-explanation"}]},{"id":"b-c2872725eba7e944d5fefea1","type":"li","scope":"retained-reading","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p58.html#b-c2872725eba7e944d5fefea1","content":"- [训练结束后，文字变成可换的候选接口](/supplement.html?return=l3-r2-p58/0#/l3-clip-use/0) 本页说明","sha256":"cdb1479f03cc11bd3e623d9342601fa3381492457d4c8c07d085cb2cc0d26923","chapter":{"id":"III","title":"视觉与语言"},"unit":{"id":"clip-pairing-to-candidates","title":"已知图文配对怎样变成可用的文字候选接口？"},"page":{"id":"clip-candidate-interface","title":"候选改变时，分数回答的问题也变了","number":1},"sectionTitle":"候选改变时，分数回答的问题也变了","currentQuestion":"从一对输入到双向损失，再到真实候选干预，哪些量由训练决定、哪些由使用者决定？","sourceSlides":["l3-r2-p58"],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-c55b13dc35e0e9b410f8b7aa","nextBlock":"b-9cd1e1c79c71805da7c519cf","prerequisites":[{"id":"b-b55fb389a499afc0cecef476","url":"#b-b55fb389a499afc0cecef476","role":"section-title"},{"id":"b-f12251795d30c9073b8e2454","url":"#b-f12251795d30c9073b8e2454","role":"problem-context"}],"relatedBlocks":[{"id":"b-ca8bf5f96c0df1f8d1cbe32a","url":"#b-ca8bf5f96c0df1f8d1cbe32a","relation":"figure-and-explanation"}]},{"id":"b-9cd1e1c79c71805da7c519cf","type":"li","scope":"retained-reading","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p58.html#b-9cd1e1c79c71805da7c519cf","content":"- [CLIP 从图文对走到候选接口｜怎样检查这项解释](/supplement.html?return=l3-r2-p58/0#/l3-clip-full-chain-evidence/0) 本页说明","sha256":"b493e1b4a74e9c4d1783e46a4cb1eedb119a9a6732f88f8d93fa34c1e1f125e3","chapter":{"id":"III","title":"视觉与语言"},"unit":{"id":"clip-pairing-to-candidates","title":"已知图文配对怎样变成可用的文字候选接口？"},"page":{"id":"clip-candidate-interface","title":"候选改变时，分数回答的问题也变了","number":1},"sectionTitle":"候选改变时，分数回答的问题也变了","currentQuestion":"从一对输入到双向损失，再到真实候选干预，哪些量由训练决定、哪些由使用者决定？","sourceSlides":["l3-r2-p58"],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-c2872725eba7e944d5fefea1","nextBlock":"b-d8de51f1feefefa81c5f2538","prerequisites":[{"id":"b-b55fb389a499afc0cecef476","url":"#b-b55fb389a499afc0cecef476","role":"section-title"},{"id":"b-f12251795d30c9073b8e2454","url":"#b-f12251795d30c9073b8e2454","role":"problem-context"}],"relatedBlocks":[{"id":"b-ca8bf5f96c0df1f8d1cbe32a","url":"#b-ca8bf5f96c0df1f8d1cbe32a","relation":"figure-and-explanation"}]},{"id":"b-d8de51f1feefefa81c5f2538","type":"li","scope":"retained-reading","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p58.html#b-d8de51f1feefefa81c5f2538","content":"- [CLIP 从图文对走到候选接口｜什么结论还不能推出](/supplement.html?return=l3-r2-p58/0#/l3-clip-full-chain-boundary/0) 本页说明","sha256":"a3b84614e2519618b7bb496b72bde7ddcc8d2ed642302e51f58dd439330916a5","chapter":{"id":"III","title":"视觉与语言"},"unit":{"id":"clip-pairing-to-candidates","title":"已知图文配对怎样变成可用的文字候选接口？"},"page":{"id":"clip-candidate-interface","title":"候选改变时，分数回答的问题也变了","number":1},"sectionTitle":"候选改变时，分数回答的问题也变了","currentQuestion":"从一对输入到双向损失，再到真实候选干预，哪些量由训练决定、哪些由使用者决定？","sourceSlides":["l3-r2-p58"],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-9cd1e1c79c71805da7c519cf","nextBlock":null,"prerequisites":[{"id":"b-b55fb389a499afc0cecef476","url":"#b-b55fb389a499afc0cecef476","role":"section-title"},{"id":"b-f12251795d30c9073b8e2454","url":"#b-f12251795d30c9073b8e2454","role":"problem-context"}],"relatedBlocks":[{"id":"b-ca8bf5f96c0df1f8d1cbe32a","url":"#b-ca8bf5f96c0df1f8d1cbe32a","relation":"figure-and-explanation"}]},{"id":"b-692ae89a502f0bf6badb092c","type":"h2","scope":"clip-lab04-candidate-controls","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p58.html#b-692ae89a502f0bf6badb092c","content":"## 在真实猫图上核对变化的量","sha256":"d539bcc5fdead43ca3a9580eb18311ece3c7510d98ea03c9f9ee02373a0161cd","chapter":{"id":"III","title":"视觉与语言"},"unit":{"id":"clip-pairing-to-candidates","title":"已知图文配对怎样变成可用的文字候选接口？"},"page":{"id":"clip-lab04-candidate-controls","title":"在真实猫图上核对变化的量","number":2},"sectionTitle":"在真实猫图上核对变化的量","currentQuestion":"从一对输入到双向损失，再到真实候选干预，哪些量由训练决定、哪些由使用者决定？","sourceSlides":["l3-r2-p58","l3-r2-p59"],"knowledgeBlock":null,"formulas":[],"previousBlock":null,"nextBlock":"b-24efd7fa00af8c66e1e50a9a","prerequisites":[{"id":"b-24efd7fa00af8c66e1e50a9a","url":"#b-24efd7fa00af8c66e1e50a9a","role":"problem-context"}],"relatedBlocks":[{"id":"b-94fb66967ed81eea63fd9be1","url":"#b-94fb66967ed81eea63fd9be1","relation":"figure-and-explanation"}]},{"id":"b-24efd7fa00af8c66e1e50a9a","type":"p","scope":"clip-lab04-candidate-controls","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p58.html#b-24efd7fa00af8c66e1e50a9a","content":"教学算例已经预测：共享图文的余弦可保持，而候选集合内的相对分数改变。","sha256":"8bd7ef97e9461420d2cafa3f45217d624050f98a53a498b584a756d60b71cde4","chapter":{"id":"III","title":"视觉与语言"},"unit":{"id":"clip-pairing-to-candidates","title":"已知图文配对怎样变成可用的文字候选接口？"},"page":{"id":"clip-lab04-candidate-controls","title":"在真实猫图上核对变化的量","number":2},"sectionTitle":"在真实猫图上核对变化的量","currentQuestion":"从一对输入到双向损失，再到真实候选干预，哪些量由训练决定、哪些由使用者决定？","sourceSlides":["l3-r2-p58","l3-r2-p59"],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-692ae89a502f0bf6badb092c","nextBlock":"b-94fb66967ed81eea63fd9be1","prerequisites":[{"id":"b-692ae89a502f0bf6badb092c","url":"#b-692ae89a502f0bf6badb092c","role":"section-title"}],"relatedBlocks":[{"id":"b-94fb66967ed81eea63fd9be1","url":"#b-94fb66967ed81eea63fd9be1","relation":"figure-and-explanation"}]},{"id":"b-94fb66967ed81eea63fd9be1","type":"figure","scope":"clip-lab04-candidate-controls","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p58.html#b-94fb66967ed81eea63fd9be1","content":"![Chelsea 猫图的已有 CLIP 四臂候选实验结果](/assets/display/fed4bc77235bae1509d5761b.webp)LAB04 已保存的真实猫图推理结果。A 为 cat/dog/horse；B 增加 tiger/fox；C 保留 A 的候选、改为 cropped 模板；D 从 B 去掉 cat。结果来自既有记录，本轮没有运行模型。","sha256":"04c5777d12b2c7948be6a306438962b1b6426b8b0d48b4f5720ebdd8fbf6f007","chapter":{"id":"III","title":"视觉与语言"},"unit":{"id":"clip-pairing-to-candidates","title":"已知图文配对怎样变成可用的文字候选接口？"},"page":{"id":"clip-lab04-candidate-controls","title":"在真实猫图上核对变化的量","number":2},"sectionTitle":"在真实猫图上核对变化的量","currentQuestion":"从一对输入到双向损失，再到真实候选干预，哪些量由训练决定、哪些由使用者决定？","sourceSlides":["l3-r2-p58","l3-r2-p59"],"knowledgeBlock":null,"formulas":[],"figure":{"caption":"LAB04 已保存的真实猫图推理结果。A 为 cat/dog/horse；B 增加 tiger/fox；C 保留 A 的候选、改为 cropped 模板；D 从 B 去掉 cat。结果来自既有记录，本轮没有运行模型。","labels":[],"description":"Chelsea 猫图的已有 CLIP 四臂候选实验结果","evidenceIdentity":"LAB04 已保存的真实猫图推理结果。A 为 cat/dog/horse；B 增加 tiger/fox；C 保留 A 的候选、改为 cropped 模板；D 从 B 去掉 cat。结果来自既有记录，本轮没有运行模型。","images":[{"src":"/assets/display/fed4bc77235bae1509d5761b.webp","alt":"Chelsea 猫图的已有 CLIP 四臂候选实验结果"}]},"previousBlock":"b-24efd7fa00af8c66e1e50a9a","nextBlock":"b-054955a55efd4f0e9f1682dc","prerequisites":[{"id":"b-692ae89a502f0bf6badb092c","url":"#b-692ae89a502f0bf6badb092c","role":"section-title"},{"id":"b-24efd7fa00af8c66e1e50a9a","url":"#b-24efd7fa00af8c66e1e50a9a","role":"problem-context"}],"relatedBlocks":[]},{"id":"b-054955a55efd4f0e9f1682dc","type":"h3","scope":"clip-real-checkpoint-evidence","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p58.html#b-054955a55efd4f0e9f1682dc","content":"### 真实猫图结果：保留模型、照片与历史身份","sha256":"8ef459f480d8633165535c202fae7d876f4ff69a4a5b1505c6fdca66a6f48ffb","chapter":{"id":"III","title":"视觉与语言"},"unit":{"id":"clip-pairing-to-candidates","title":"已知图文配对怎样变成可用的文字候选接口？"},"page":{"id":"clip-lab04-candidate-controls","title":"在真实猫图上核对变化的量","number":2},"sectionTitle":"真实猫图结果：保留模型、照片与历史身份","currentQuestion":"从一对输入到双向损失，再到真实候选干预，哪些量由训练决定、哪些由使用者决定？","sourceSlides":["l3-r2-p58","l3-r2-p59"],"knowledgeBlock":"clip-real-checkpoint-evidence","formulas":[],"previousBlock":"b-94fb66967ed81eea63fd9be1","nextBlock":"b-39665916cb74ce1310551303","prerequisites":[{"id":"b-692ae89a502f0bf6badb092c","url":"#b-692ae89a502f0bf6badb092c","role":"section-title"},{"id":"b-24efd7fa00af8c66e1e50a9a","url":"#b-24efd7fa00af8c66e1e50a9a","role":"problem-context"}],"relatedBlocks":[{"id":"b-94fb66967ed81eea63fd9be1","url":"#b-94fb66967ed81eea63fd9be1","relation":"figure-and-explanation"},{"id":"b-39665916cb74ce1310551303","url":"#b-39665916cb74ce1310551303","relation":"same-reasoning"},{"id":"b-f7420cc532f6a99410b5a9ae","url":"#b-f7420cc532f6a99410b5a9ae","relation":"same-reasoning"},{"id":"b-59bc4d9eddd71dc79ae1aed3","url":"#b-59bc4d9eddd71dc79ae1aed3","relation":"same-reasoning"},{"id":"b-64711d8f36a2686470a84ea7","url":"#b-64711d8f36a2686470a84ea7","relation":"same-reasoning"}]},{"id":"b-39665916cb74ce1310551303","type":"p","scope":"clip-real-checkpoint-evidence","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p58.html#b-39665916cb74ce1310551303","content":"当前 LAB04 保存真实 Chelsea 猫照片与 OpenAI CLIP ViT-B/32 实验。2026-09-22 的本地复核使用固定 checkpoint revision `3d74acf9a28c67741b2f4f2ea7635f0aaf6f0268`，没有重新训练 CLIP；当前整理也没有重跑该推理。图片先转 RGB、缩放并中心裁剪到 $224\\times224$，再按模型配置归一化。","sha256":"08592cac5e85947afd40374d30d6030489704a2787422ccca7752470cb945f62","chapter":{"id":"III","title":"视觉与语言"},"unit":{"id":"clip-pairing-to-candidates","title":"已知图文配对怎样变成可用的文字候选接口？"},"page":{"id":"clip-lab04-candidate-controls","title":"在真实猫图上核对变化的量","number":2},"sectionTitle":"真实猫图结果：保留模型、照片与历史身份","currentQuestion":"从一对输入到双向损失，再到真实候选干预，哪些量由训练决定、哪些由使用者决定？","sourceSlides":["l3-r2-p58","l3-r2-p59"],"knowledgeBlock":"clip-real-checkpoint-evidence","formulas":[{"latex":"224\\times224","definitionContext":"当前 LAB04 保存真实 Chelsea 猫照片与 OpenAI CLIP ViT-B/32 实验。2026-09-22 的本地复核使用固定 checkpoint revision `3d74acf9a28c67741b2f4f2ea7635f0aaf6f0268`，没有重新训练 CLIP；当前整理也没有重跑该推理。图片先转 RGB、缩放并中心裁剪到 $224\\times224$，再按模型配置归一化。\n基础 cat/dog/horse 候选下，cat 余弦约 0.28691、候选 softmax 约 0.99577；添加 tiger 与 fox 后，共享 cat 的余弦保持不变，softmax 降至约 0.98195；再删除 cat，tiger 成为剩余候选的最高项。新旧猫图余弦最大差约 $1.04\\times10^{-7}$，把后来的复核与历史材料连接起来。\n正面结果是模型可以在这张真实照片上把图像与合理文字对应，并且其候选接口的变化符合预期。数值 0.99577 是给定候选集合内的相对分配，不是“现实中有猫的校准概率”。一张猫图也不能替代数据集级泛化评测。\nLAB04 负责模型输入、候选、模板和关系检查；LAB12 复用同一模型产生的完整 512 维向量，研究图库缓存与搜索。不要把四个候选分数拼成所谓图像表示；真正可复用的是编码器输出的向量。先在 LAB04 理解分数身份，再在 LAB12 看系统组织，就无需重复训练或建立两份同义实验。"}],"previousBlock":"b-054955a55efd4f0e9f1682dc","nextBlock":"b-f7420cc532f6a99410b5a9ae","prerequisites":[{"id":"b-054955a55efd4f0e9f1682dc","url":"#b-054955a55efd4f0e9f1682dc","role":"section-title"},{"id":"b-24efd7fa00af8c66e1e50a9a","url":"#b-24efd7fa00af8c66e1e50a9a","role":"problem-context"}],"relatedBlocks":[{"id":"b-94fb66967ed81eea63fd9be1","url":"#b-94fb66967ed81eea63fd9be1","relation":"figure-and-explanation"},{"id":"b-054955a55efd4f0e9f1682dc","url":"#b-054955a55efd4f0e9f1682dc","relation":"same-reasoning"},{"id":"b-f7420cc532f6a99410b5a9ae","url":"#b-f7420cc532f6a99410b5a9ae","relation":"same-reasoning"},{"id":"b-59bc4d9eddd71dc79ae1aed3","url":"#b-59bc4d9eddd71dc79ae1aed3","relation":"same-reasoning"},{"id":"b-64711d8f36a2686470a84ea7","url":"#b-64711d8f36a2686470a84ea7","relation":"same-reasoning"}]},{"id":"b-f7420cc532f6a99410b5a9ae","type":"p","scope":"clip-real-checkpoint-evidence","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p58.html#b-f7420cc532f6a99410b5a9ae","content":"基础 cat/dog/horse 候选下，cat 余弦约 0.28691、候选 softmax 约 0.99577；添加 tiger 与 fox 后，共享 cat 的余弦保持不变，softmax 降至约 0.98195；再删除 cat，tiger 成为剩余候选的最高项。新旧猫图余弦最大差约 $1.04\\times10^{-7}$，把后来的复核与历史材料连接起来。","sha256":"c7061a8ad0794772fc5fd4b5f0bdb6e66e098c0aaa2732e4c041ff8bb739810e","chapter":{"id":"III","title":"视觉与语言"},"unit":{"id":"clip-pairing-to-candidates","title":"已知图文配对怎样变成可用的文字候选接口？"},"page":{"id":"clip-lab04-candidate-controls","title":"在真实猫图上核对变化的量","number":2},"sectionTitle":"真实猫图结果：保留模型、照片与历史身份","currentQuestion":"从一对输入到双向损失，再到真实候选干预，哪些量由训练决定、哪些由使用者决定？","sourceSlides":["l3-r2-p58","l3-r2-p59"],"knowledgeBlock":"clip-real-checkpoint-evidence","formulas":[{"latex":"1.04\\times10^{-7}","definitionContext":"当前 LAB04 保存真实 Chelsea 猫照片与 OpenAI CLIP ViT-B/32 实验。2026-09-22 的本地复核使用固定 checkpoint revision `3d74acf9a28c67741b2f4f2ea7635f0aaf6f0268`，没有重新训练 CLIP；当前整理也没有重跑该推理。图片先转 RGB、缩放并中心裁剪到 $224\\times224$，再按模型配置归一化。\n基础 cat/dog/horse 候选下，cat 余弦约 0.28691、候选 softmax 约 0.99577；添加 tiger 与 fox 后，共享 cat 的余弦保持不变，softmax 降至约 0.98195；再删除 cat，tiger 成为剩余候选的最高项。新旧猫图余弦最大差约 $1.04\\times10^{-7}$，把后来的复核与历史材料连接起来。\n正面结果是模型可以在这张真实照片上把图像与合理文字对应，并且其候选接口的变化符合预期。数值 0.99577 是给定候选集合内的相对分配，不是“现实中有猫的校准概率”。一张猫图也不能替代数据集级泛化评测。\nLAB04 负责模型输入、候选、模板和关系检查；LAB12 复用同一模型产生的完整 512 维向量，研究图库缓存与搜索。不要把四个候选分数拼成所谓图像表示；真正可复用的是编码器输出的向量。先在 LAB04 理解分数身份，再在 LAB12 看系统组织，就无需重复训练或建立两份同义实验。"}],"previousBlock":"b-39665916cb74ce1310551303","nextBlock":"b-59bc4d9eddd71dc79ae1aed3","prerequisites":[{"id":"b-054955a55efd4f0e9f1682dc","url":"#b-054955a55efd4f0e9f1682dc","role":"section-title"},{"id":"b-24efd7fa00af8c66e1e50a9a","url":"#b-24efd7fa00af8c66e1e50a9a","role":"problem-context"}],"relatedBlocks":[{"id":"b-94fb66967ed81eea63fd9be1","url":"#b-94fb66967ed81eea63fd9be1","relation":"figure-and-explanation"},{"id":"b-054955a55efd4f0e9f1682dc","url":"#b-054955a55efd4f0e9f1682dc","relation":"same-reasoning"},{"id":"b-39665916cb74ce1310551303","url":"#b-39665916cb74ce1310551303","relation":"same-reasoning"},{"id":"b-59bc4d9eddd71dc79ae1aed3","url":"#b-59bc4d9eddd71dc79ae1aed3","relation":"same-reasoning"},{"id":"b-64711d8f36a2686470a84ea7","url":"#b-64711d8f36a2686470a84ea7","relation":"same-reasoning"}]},{"id":"b-59bc4d9eddd71dc79ae1aed3","type":"p","scope":"clip-real-checkpoint-evidence","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p58.html#b-59bc4d9eddd71dc79ae1aed3","content":"正面结果是模型可以在这张真实照片上把图像与合理文字对应，并且其候选接口的变化符合预期。数值 0.99577 是给定候选集合内的相对分配，不是“现实中有猫的校准概率”。一张猫图也不能替代数据集级泛化评测。","sha256":"1fb9bd54f6a6d811bd1b68919726fd1177e6a5264ddb1afa54180876084827ba","chapter":{"id":"III","title":"视觉与语言"},"unit":{"id":"clip-pairing-to-candidates","title":"已知图文配对怎样变成可用的文字候选接口？"},"page":{"id":"clip-lab04-candidate-controls","title":"在真实猫图上核对变化的量","number":2},"sectionTitle":"真实猫图结果：保留模型、照片与历史身份","currentQuestion":"从一对输入到双向损失，再到真实候选干预，哪些量由训练决定、哪些由使用者决定？","sourceSlides":["l3-r2-p58","l3-r2-p59"],"knowledgeBlock":"clip-real-checkpoint-evidence","formulas":[],"previousBlock":"b-f7420cc532f6a99410b5a9ae","nextBlock":"b-64711d8f36a2686470a84ea7","prerequisites":[{"id":"b-054955a55efd4f0e9f1682dc","url":"#b-054955a55efd4f0e9f1682dc","role":"section-title"},{"id":"b-24efd7fa00af8c66e1e50a9a","url":"#b-24efd7fa00af8c66e1e50a9a","role":"problem-context"}],"relatedBlocks":[{"id":"b-94fb66967ed81eea63fd9be1","url":"#b-94fb66967ed81eea63fd9be1","relation":"figure-and-explanation"},{"id":"b-054955a55efd4f0e9f1682dc","url":"#b-054955a55efd4f0e9f1682dc","relation":"same-reasoning"},{"id":"b-39665916cb74ce1310551303","url":"#b-39665916cb74ce1310551303","relation":"same-reasoning"},{"id":"b-f7420cc532f6a99410b5a9ae","url":"#b-f7420cc532f6a99410b5a9ae","relation":"same-reasoning"},{"id":"b-64711d8f36a2686470a84ea7","url":"#b-64711d8f36a2686470a84ea7","relation":"same-reasoning"}]},{"id":"b-64711d8f36a2686470a84ea7","type":"p","scope":"clip-real-checkpoint-evidence","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p58.html#b-64711d8f36a2686470a84ea7","content":"LAB04 负责模型输入、候选、模板和关系检查；LAB12 复用同一模型产生的完整 512 维向量，研究图库缓存与搜索。不要把四个候选分数拼成所谓图像表示；真正可复用的是编码器输出的向量。先在 LAB04 理解分数身份，再在 LAB12 看系统组织，就无需重复训练或建立两份同义实验。","sha256":"3fd55da4e643785a3ea3b5949b03d73e798ccf3ae1b5812997a45c182c9d7898","chapter":{"id":"III","title":"视觉与语言"},"unit":{"id":"clip-pairing-to-candidates","title":"已知图文配对怎样变成可用的文字候选接口？"},"page":{"id":"clip-lab04-candidate-controls","title":"在真实猫图上核对变化的量","number":2},"sectionTitle":"真实猫图结果：保留模型、照片与历史身份","currentQuestion":"从一对输入到双向损失，再到真实候选干预，哪些量由训练决定、哪些由使用者决定？","sourceSlides":["l3-r2-p58","l3-r2-p59"],"knowledgeBlock":"clip-real-checkpoint-evidence","formulas":[],"previousBlock":"b-59bc4d9eddd71dc79ae1aed3","nextBlock":"b-48a018c1cdfb14b237f9b6cb","prerequisites":[{"id":"b-054955a55efd4f0e9f1682dc","url":"#b-054955a55efd4f0e9f1682dc","role":"section-title"},{"id":"b-24efd7fa00af8c66e1e50a9a","url":"#b-24efd7fa00af8c66e1e50a9a","role":"problem-context"}],"relatedBlocks":[{"id":"b-94fb66967ed81eea63fd9be1","url":"#b-94fb66967ed81eea63fd9be1","relation":"figure-and-explanation"},{"id":"b-054955a55efd4f0e9f1682dc","url":"#b-054955a55efd4f0e9f1682dc","relation":"same-reasoning"},{"id":"b-39665916cb74ce1310551303","url":"#b-39665916cb74ce1310551303","relation":"same-reasoning"},{"id":"b-f7420cc532f6a99410b5a9ae","url":"#b-f7420cc532f6a99410b5a9ae","relation":"same-reasoning"},{"id":"b-59bc4d9eddd71dc79ae1aed3","url":"#b-59bc4d9eddd71dc79ae1aed3","relation":"same-reasoning"}]},{"id":"b-48a018c1cdfb14b237f9b6cb","type":"h3","scope":"dialogue","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p58.html#b-48a018c1cdfb14b237f9b6cb","content":"### 沿着图解追问","sha256":"f16be0424762a44344338696f06ef266c4571d15367ae2cbbd28fac98ac7e9cb","chapter":{"id":"III","title":"视觉与语言"},"unit":{"id":"clip-pairing-to-candidates","title":"已知图文配对怎样变成可用的文字候选接口？"},"page":{"id":"clip-lab04-candidate-controls","title":"在真实猫图上核对变化的量","number":2},"sectionTitle":"在真实猫图上核对变化的量","currentQuestion":"从一对输入到双向损失，再到真实候选干预，哪些量由训练决定、哪些由使用者决定？","sourceSlides":["l3-r2-p58","l3-r2-p59"],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-64711d8f36a2686470a84ea7","nextBlock":"b-feec39ed6d00fa6ad145a1f8","prerequisites":[{"id":"b-054955a55efd4f0e9f1682dc","url":"#b-054955a55efd4f0e9f1682dc","role":"section-title"},{"id":"b-24efd7fa00af8c66e1e50a9a","url":"#b-24efd7fa00af8c66e1e50a9a","role":"problem-context"}],"relatedBlocks":[{"id":"b-94fb66967ed81eea63fd9be1","url":"#b-94fb66967ed81eea63fd9be1","relation":"figure-and-explanation"}]},{"id":"b-feec39ed6d00fa6ad145a1f8","type":"p","scope":"dialogue","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p58.html#b-feec39ed6d00fa6ad145a1f8","content":"以下是备课用的模拟师生对话，不是课堂实录。","sha256":"b25685f04cd1b885f3a5999e6a912460c7424021cdc5b2190584242086a82681","chapter":{"id":"III","title":"视觉与语言"},"unit":{"id":"clip-pairing-to-candidates","title":"已知图文配对怎样变成可用的文字候选接口？"},"page":{"id":"clip-lab04-candidate-controls","title":"在真实猫图上核对变化的量","number":2},"sectionTitle":"在真实猫图上核对变化的量","currentQuestion":"从一对输入到双向损失，再到真实候选干预，哪些量由训练决定、哪些由使用者决定？","sourceSlides":["l3-r2-p58","l3-r2-p59"],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-48a018c1cdfb14b237f9b6cb","nextBlock":"b-68dac1c085af6e65f27c5515","prerequisites":[{"id":"b-48a018c1cdfb14b237f9b6cb","url":"#b-48a018c1cdfb14b237f9b6cb","role":"section-title"},{"id":"b-24efd7fa00af8c66e1e50a9a","url":"#b-24efd7fa00af8c66e1e50a9a","role":"problem-context"}],"relatedBlocks":[{"id":"b-94fb66967ed81eea63fd9be1","url":"#b-94fb66967ed81eea63fd9be1","relation":"figure-and-explanation"}]},{"id":"b-68dac1c085af6e65f27c5515","type":"p","scope":"dialogue","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p58.html#b-68dac1c085af6e65f27c5515","content":"**学生：**增加一个候选，怎么连猫的概率都变了？","sha256":"491fc9e14f1912ea4bbbc4e9d1c5a17b30cd6ae22d093d54b6b170c9432077c8","chapter":{"id":"III","title":"视觉与语言"},"unit":{"id":"clip-pairing-to-candidates","title":"已知图文配对怎样变成可用的文字候选接口？"},"page":{"id":"clip-lab04-candidate-controls","title":"在真实猫图上核对变化的量","number":2},"sectionTitle":"在真实猫图上核对变化的量","currentQuestion":"从一对输入到双向损失，再到真实候选干预，哪些量由训练决定、哪些由使用者决定？","sourceSlides":["l3-r2-p58","l3-r2-p59"],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-feec39ed6d00fa6ad145a1f8","nextBlock":"b-6613a34dadedb0edcb265e6a","prerequisites":[{"id":"b-48a018c1cdfb14b237f9b6cb","url":"#b-48a018c1cdfb14b237f9b6cb","role":"section-title"},{"id":"b-24efd7fa00af8c66e1e50a9a","url":"#b-24efd7fa00af8c66e1e50a9a","role":"problem-context"}],"relatedBlocks":[{"id":"b-94fb66967ed81eea63fd9be1","url":"#b-94fb66967ed81eea63fd9be1","relation":"figure-and-explanation"}]},{"id":"b-6613a34dadedb0edcb265e6a","type":"p","scope":"dialogue","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p58.html#b-6613a34dadedb0edcb265e6a","content":"**教师：**余弦没有变；概率分母多了一个竞争项。它回答的是这组选项里的相对分配。请同时保存原始相似度和候选集合。","sha256":"335d4c8c80f3ea1e4a56444e3d790a917f3d860fb0499b78a839f28a7ab795d6","chapter":{"id":"III","title":"视觉与语言"},"unit":{"id":"clip-pairing-to-candidates","title":"已知图文配对怎样变成可用的文字候选接口？"},"page":{"id":"clip-lab04-candidate-controls","title":"在真实猫图上核对变化的量","number":2},"sectionTitle":"在真实猫图上核对变化的量","currentQuestion":"从一对输入到双向损失，再到真实候选干预，哪些量由训练决定、哪些由使用者决定？","sourceSlides":["l3-r2-p58","l3-r2-p59"],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-68dac1c085af6e65f27c5515","nextBlock":"b-ad2aabd871e0840c1a9b6d09","prerequisites":[{"id":"b-48a018c1cdfb14b237f9b6cb","url":"#b-48a018c1cdfb14b237f9b6cb","role":"section-title"},{"id":"b-24efd7fa00af8c66e1e50a9a","url":"#b-24efd7fa00af8c66e1e50a9a","role":"problem-context"}],"relatedBlocks":[{"id":"b-94fb66967ed81eea63fd9be1","url":"#b-94fb66967ed81eea63fd9be1","relation":"figure-and-explanation"}]},{"id":"b-ad2aabd871e0840c1a9b6d09","type":"h2","scope":"practice","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p58.html#b-ad2aabd871e0840c1a9b6d09","content":"## 沿这个问题继续","sha256":"cf44fcc62782fb4f6bdc42c16c5187d57556a03cb7ce6065e7bd9c36d8e7d733","chapter":{"id":"III","title":"视觉与语言"},"unit":{"id":"clip-pairing-to-candidates","title":"已知图文配对怎样变成可用的文字候选接口？"},"page":{"id":"clip-lab04-candidate-controls","title":"在真实猫图上核对变化的量","number":2},"sectionTitle":"在真实猫图上核对变化的量","currentQuestion":"从一对输入到双向损失，再到真实候选干预，哪些量由训练决定、哪些由使用者决定？","sourceSlides":["l3-r2-p58","l3-r2-p59"],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-6613a34dadedb0edcb265e6a","nextBlock":"b-196ea3d15b85ad404d6ec034","prerequisites":[{"id":"b-48a018c1cdfb14b237f9b6cb","url":"#b-48a018c1cdfb14b237f9b6cb","role":"section-title"},{"id":"b-24efd7fa00af8c66e1e50a9a","url":"#b-24efd7fa00af8c66e1e50a9a","role":"problem-context"}],"relatedBlocks":[{"id":"b-94fb66967ed81eea63fd9be1","url":"#b-94fb66967ed81eea63fd9be1","relation":"figure-and-explanation"}]},{"id":"b-196ea3d15b85ad404d6ec034","type":"p","scope":"practice","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p58.html#b-196ea3d15b85ad404d6ec034","content":"先用给定 logits 复算两组 softmax，再到 LAB04 对照真实猫图的四臂结果。分别记录余弦与候选分数；需要研究图库搜索时，再进入 LAB12。","sha256":"b8695bb1b1bba8ad0c7695972964a508c00ec40840e1197f42629041cad9e7e4","chapter":{"id":"III","title":"视觉与语言"},"unit":{"id":"clip-pairing-to-candidates","title":"已知图文配对怎样变成可用的文字候选接口？"},"page":{"id":"clip-lab04-candidate-controls","title":"在真实猫图上核对变化的量","number":2},"sectionTitle":"在真实猫图上核对变化的量","currentQuestion":"从一对输入到双向损失，再到真实候选干预，哪些量由训练决定、哪些由使用者决定？","sourceSlides":["l3-r2-p58","l3-r2-p59"],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-ad2aabd871e0840c1a9b6d09","nextBlock":"b-020ee6e9d1f2bd96e7d874d3","prerequisites":[{"id":"b-ad2aabd871e0840c1a9b6d09","url":"#b-ad2aabd871e0840c1a9b6d09","role":"section-title"},{"id":"b-24efd7fa00af8c66e1e50a9a","url":"#b-24efd7fa00af8c66e1e50a9a","role":"problem-context"}],"relatedBlocks":[{"id":"b-94fb66967ed81eea63fd9be1","url":"#b-94fb66967ed81eea63fd9be1","relation":"figure-and-explanation"}]},{"id":"b-020ee6e9d1f2bd96e7d874d3","type":"h4","scope":"practice","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p58.html#b-020ee6e9d1f2bd96e7d874d3","content":"#### 按问题回查的讲解与短练习（1）","sha256":"0480869f0a66009b015a757d67aed6061011e378987f98a024f4e8bfc5386152","chapter":{"id":"III","title":"视觉与语言"},"unit":{"id":"clip-pairing-to-candidates","title":"已知图文配对怎样变成可用的文字候选接口？"},"page":{"id":"clip-lab04-candidate-controls","title":"在真实猫图上核对变化的量","number":2},"sectionTitle":"在真实猫图上核对变化的量","currentQuestion":"从一对输入到双向损失，再到真实候选干预，哪些量由训练决定、哪些由使用者决定？","sourceSlides":["l3-r2-p58","l3-r2-p59"],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-196ea3d15b85ad404d6ec034","nextBlock":null,"prerequisites":[{"id":"b-ad2aabd871e0840c1a9b6d09","url":"#b-ad2aabd871e0840c1a9b6d09","role":"section-title"},{"id":"b-24efd7fa00af8c66e1e50a9a","url":"#b-24efd7fa00af8c66e1e50a9a","role":"problem-context"}],"relatedBlocks":[{"id":"b-94fb66967ed81eea63fd9be1","url":"#b-94fb66967ed81eea63fd9be1","relation":"figure-and-explanation"}]}]}